基于GNSS融合的单目视觉SLAM算法

Monocular Vision SLAM Algorithm Based on GNSS Fusion

针对单目视觉SLAM(Simultaneous Localization and Mapping,同时定位与建图)算法无法恢复地理尺度信息、位姿估计误差累积大等问题,提出了一种基于GNSS融合的单目视觉SLAM算法。该算法基于图优化理论,在视觉里程计中构建相似变换估计模型求解真实尺度,在后端优化中加入GNSS全局地理位置节点,设计优化求解策略,提高关键帧位姿以及三维地图点位置估计的准确性,最后在序列末尾对全局地图离线迭代优化,保证地图构建的全局一致性。实验表明,在室外环境下,运动轨迹在5 km以内时,所提出方法的尺度估计误差在0.1%以内,轨迹误差在2 m以内,满足实际应用需求。

Aiming at the problems that the monocular visual SLAM(Simultaneous Localization and Mapping)algorithm could not recover the geographic scale information and the accumulation of pose estimation errors is large,a monocular vision SLAM algorithm based on GNSS fusion has been proposed.Based on the graph optimization theory,the algorithm built a similarity transformation estimation model in the visual odometry to solve the real scale,added the GNSS global location node in the back-end optimization,designed the optimization solution strategy,and improved the key frame pose and 3D map point position estimation.Finally,at the end of the sequence,the global map was iteratively optimized offline to ensure the global consistency of map construction.Experiments showed that in the outdoor environment,when the motion trajectory was within 5 km,the scale estimation error of the method proposed in this paper was within 0.1%,and the trajectory error was within 2 m,which could meet the needs of practical applications.